normalblockr

• noise_covariance gains a "full" shape for this family, whose Sigma is the p x p residual covariance rather than a diagonal noise term. "diagonal" is the default: a full Sigma spends p(p+1)/2 parameters that drown the mean structure the criteria are weighing, and selects the number of clusters markedly worse as p approaches n. Asking for sparsity > 0 implies "full", since a diagonal precision matrix has nothing for the graphical lasso to penalize.
• normal_block_sequential() chains the two families: a mean-block fit, then a variance-block fit on its residuals. A heuristic two-stage estimator, not a joint model.
• NormalBlockData$new() gains a zeros argument, to carry an explicit zero-inflation mask when the matrix handed to a model is no longer the one carrying the zeros (the residuals of a first stage, say).

The graphical lasso is now in-package

This also fixes the segfault CRAN reported on

r-devel-linux-x86_64-fedora-gcc when re-building

breast-cancer-proteomics.Rmd (address 0x580, cause 'memory not mapped'),

which our own CI had been hitting intermittently on Linux.

• glassoFast is no longer a dependency. The C++ (V)EM used to call back into R once per M-step to reach it -- thousands of round trips from inside a single .Call for one sparsity path, and a recurring source of intermittent crashes. The solver now lives in src/graphical_lasso.h.
• Two bugs in glassoFast are fixed on the way. When the empirical covariance has no off-diagonal mass the problem separates and the solution is diagonal; it returned 1 / max(rho_ii, eps) there, dropping the variance term, so with an unpenalized diagonal it returned about 9e15 instead of 1 / S_ii -- and every q = 1 problem takes that branch. Separately, its inner loop could not terminate on non-finite input.
• Each M-step warm-starts from the previous one, at a tightened threshold: faster *and* closer to the optimum than the previous cold starts.

Bug fixes

• Zero-inflation with more than one zero-inflation covariate (X0 with two or more columns) failed with "non-conformable arguments": B0 was built transposed. A second bug behind it made kappa non-finite when the response had no zeros at all. Both were unreachable before, so no working fit changes.
• Fixed two crashes on the penalized path, one of them a long-standing memory-safety bug (an R object cached in a C++ static).

Other changes

• A collection over cluster counts no longer repeats work that does not depend on the cluster count: the OLS fit, the zero-inflation logistic regressions, and the shareable part of each clustering heuristic (one hierarchical tree cut at each q, one eigendecomposition, one lossless row compression) are computed once. Fitted values are unchanged.
• Printed model descriptions now name the residual-covariance shape for every family, and the zero-inflated variance-block models name their family.